Multi-task Scale Adaptive Ladder Network for Crowd Counting

As the population increases, problems such as crowds and traffic jams have emerged one after another. How to effectively achieve accurate human flow monitoring has become an urgent problem of today's society. This paper proposes a multi-task scale adaptive ladder network (MT-SALN) for generating high-accuracy crowd density maps. This network, based on VGG-16 network, consists of several sets of Adaptive Dilated-Convolution Module (ADCM), a Position Recalibration Branch (PRB) and a Density Estimation Branch (DEB). We employ ADCM in different stages to broaden the width of the network and introduce weights for each channel parameter through an attention mechanism. The residual structure enables the network model to have a back propagation ability even though the number of network layers is large. In addition, transposed convolution is used to upsample the features so that they can be merged with other layers' features to generate a more refined density map with high resolution. The existence of PRB can effectively guide the network to generate crowd density at the correct location and accelerate network convergence. The ladder architecture is beneficial to produce high-quality density maps. Extensive experiments on challenging crowd counting datasets (UCF_CC_50, Shanghaitech) demonstrate the effectiveness of the proposed approach.